Email campaign attribution

The question is "which campaigns made money?". Without a holdout group you can't answer it causally. You can answer "which campaigns were the last thing a buyer engaged with before ordering", and say so on every output. Done carefully, that is still the most useful ranking marketing has.

Pipeline

extract engagement -> extract orders -> resolve identity -> attribute in a window -> rank -> report with context

Read-only against every source system. Land raw extracts first (files or a local database such as DuckDB or SQLite), then build the model from them, so a rerun doesn't call the APIs again.

Extract